Researchers introduce ColorBrowserAgent, a knowledge-evolving agent designed to address site heterogeneity and long-horizon instability in autonomous web automation. The system utilizes human-in-the-loop knowledge adaptation to convert sparse feedback into reusable domain knowledge and employs progressive summarization to stabilize extended interactions.
- Achieves a state-of-the-art success rate of 71.2% on WebArena.
- Maintains 47.4% performance under zero-shot transfer settings on WebChoreArena.
- Improves user satisfaction by 19.3% relatively in commercial deployment.
The approach demonstrates robustness in real-world scenarios by consistently outperforming strong baselines across extensive experiments and industrial applications.